Moonshot opened Kimi K3.1 Pro weights (2.8T MoE, 104B active, 1M context); Ricursive released the dense, Apache-2.0 Ripple-65B Preview with 80.5 on TerminalBench 2.1. "You don't need long context": frequent compaction keeps accuracy on long tasks with short windows, questioning the need for ultra-long-context models. Reasoning efficiency: LoCo teaches models when to stop thinking, cutting math completions by 41.7–45.4% without losing accuracy. Agents: Axion 1.0 releases a fully open agent foundation-model recipe; ToolSpace synthesizes tool environments from MCP execution traces.
Large Language Models
1
Think Less, Locally: Local Preference Learning for Reasoning Compression
Large reasoning models often keep deliberating after they already have enough information. LoCo locates candidate overthinking points in the model's own traces, freezes the prefix, regenerates shorter verified continuations and trains with DPO on the resulting pairs. Trained only on self-generated math traces, it shortens completions by 41.7–45.4% on five math benchmarks and 23.4–36.3% on three code benchmarks with accuracy maintained or improved.
Why it mattersReasoning cost is now a major serving bottleneck, and LoCo targets the local stop decision rather than whole-trajectory length penalties. The transfer from math-only training to code suggests the model learns a general sense of when to conclude.
Later tokens cost more to generate than earlier ones, by up to about 7x on frontier open-weight models. With frequent compaction, models lose little or no accuracy on long-horizon and long-input tasks while keeping contexts short, and long-horizon software engineering holds up even with one action per context. The authors propose semantic state space models (SSSMs) and question whether long-context LLMs are necessary.
Why it mattersIt directly challenges the race toward million-token windows and argues the cheaper path is better context management. If it holds up, agent harnesses may invest in compaction rather than raw context length.
Route Anything: A Study of Model and Policy Routing for General-Purpose AI Systems
The paper frames general-purpose AI as orchestration of many specialized models, tools and policies, and gives a unified framework plus theory for quality-cost routing. It builds an omni-modal trajectory matrix of 25 benchmarks with 374,401 scored runs from 452 candidate configurations. Per-input oracle routing beats the best fixed candidate by 5.4 points, and learned routers improve efficiency by 8.82x on average versus the reference model.
Why it mattersThe released-scale routing matrix is a useful public resource for router research across modalities. The gap between oracle and learned routers also quantifies how much headroom remains.
Axion 1.0: Open Agent Foundation Models, Simplified
Axion 1.0 is a fully open agent foundation model trained with two lightweight SFT stages followed by RL. Stage one builds self-contained reasoning from synthesized, adversarially strengthened problems; stage two adds tool-dependent reasoning from automatically verifiable tasks. The authors report strong results across agentic benchmarks, competitive with leading commercial models on several, and release the full data, prompts and training code.
Why it mattersFully reproducible agent training recipes are still rare, so the release lowers the barrier for academic labs. The emphasis on simplicity makes it a useful baseline to compare more complex agent pipelines against.
ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents
ToolSpace builds training environments for MCP tool-using agents by starting from tool interactions in execution traces, mining dependencies deterministically and clustering them into environments. It generates state-aware task chains with assertion checks and keeps only those that replay successfully. It yields harder benchmark variants and improves Qwen3.6-35B-A3B and GLM-5.2 across seven long-horizon tool-use benchmarks.
Why it mattersEnvironment scarcity is a key limit on agent RL, and grounding synthesis in real execution traces makes tasks verifiable by construction. It continues the same group's push (after CoworkOS) toward scalable agent training infrastructure.
Adversarial Speculative Reasoning Improves Long Horizon Agents
The paper argues speculative decoding affects long-horizon agent behavior, not just speed: numerical noise grows within verification windows and can shift outputs over many steps. It trains a draft model to diverge from the target at critical reasoning points so key tokens come from the target under lower noise. This raises end-to-end success on agentic coding benchmarks without changing target weights, using draft components of only 0.06–0.28% of hosted parameters.
Why it mattersIt reframes an inference optimization as a lever on agent quality, which matters for anyone serving open models in agent loops. The tiny parameter overhead makes it cheap to try in production.
Moonshot published open weights for Kimi K3.1 Pro on Hugging Face after its earlier app and API launch. It is a natively multimodal 2.8T-parameter MoE with 104B active, 896 experts, ~1M context, hybrid KDA/Gated DSA attention and MXFP4 quantization-aware weights. Self-reported results include 65.7 on Terminal-Bench 4.0 (Claude Opus 5.5: 66.4), 92.1 on BrowseComp and 79.6 on OSWorld V2.1, under a custom Kimi K3.1 Pro License.
Why it mattersIt is among the largest open-weight agentic models and sits close to frontier proprietary models on several agent benchmarks. Native MXFP4 weights also make a 3T-class model more practical to self-host.
Ricursive Intelligence released Ripple-65B Preview, a general-purpose variant of its chip-design model, under Apache 2.0. It is a dense 65B transformer pretrained on 12T tokens with up to 1M context. The model card reports 80.5 on TerminalBench 2.1, 76.5 on SWE-bench Verified and 85.2 on GPQA Diamond, with total R&D cost cited at about $32M; NVFP4, base and DFlash2 variants are also published.
Why it mattersA permissively licensed dense 65B model claiming coding results on par with much larger MoEs is notable for teams that want simpler deployment. The claimed cost is a data point on how cheap competitive open models have become, though results are self-reported.
GigaSpeech 3: Scaling Multi-Dialect Arabic Speech Recognition with Weakly Labeled Data
GigaSpeech 3 is a ~39,000-hour, 10M-segment weakly labeled corpus for five Arabic dialects, bootstrapped from 60 hours of manual transcripts per dialect with noisy-student refinement. Dialect-specific models reach 23.86% macro-average WER on GigaSpeechBench, a 16.6% relative reduction over the strongest third-party system. A Zipformer trained on it had the lowest WER of six models in all eight external test conditions, and the corpus and recipes are released.
Why it mattersDialectal Arabic remains badly underserved by ASR, and this is a large open resource for it. The weak-labeling recipe is also a template for other low-resource dialect families.
InteractionBench: Measuring Timing, Task Accuracy, Whisper and Volume, and Grounding in Real-Time Voice Systems
InteractionBench evaluates real-time voice systems on timing, spoken-task accuracy, whisper/volume control and groundedness over 20+ turns, using the same recorded user audio. It runs 221 items across 13 scenarios and 1,519 graded evaluations on GPT-Live, GPT-Realtime, Gemini Live, Moshi, PersonaPlex, Freeze-Omni and a turn-based cascade. The cascade scored best on tasks (38/50) but second-lowest on timing, and hosted systems passed 67–88% of groundedness probes versus 14–35% for open full-duplex models.
Why it mattersIt is one of the few benchmarks that measures timing and task quality together on identical audio, exposing the trade-off between them. The large groundedness gap shows where open full-duplex models most need work.
Voice-Native LLMs Hear Frustration but Do Not Listen to It
Using Brunswik's lens model and a 2x2 design crossing frustrated wording with frustrated voice over 353 tasks, the paper tests whether voice-native LLMs use vocal frustration. With neutral wording, all 8 models label frustrated voices as frustrated in at most 3.1% of clips, though 7 detect acted anger in 73–100%. Linear probes still decode frustration from the audio encoders of the two open-weight models (0.67 and 0.69), and the 1,412-clip VoxAffect dataset is released.
Why it mattersFor voice agents in customer support, missing a frustrated caller is a real failure even when the task succeeds. The paper warns that unchanged accuracy is not evidence of robustness unless perception is checked.
